Efficient MFCC Computation for Keyword Spotting on Embedded Systems: Optimization Techniques and Performance Analysis

Recent advancements in neural networks for edge devices, like model quantization and quantization-aware training (QAT), have significantly improved Keyword Spotting (KWS) performance. However, the audio preprocessing stage, especially Mel-frequency Cepstral Coefficient (MFCC) feature extraction, still poses challenges. This paper aims to optimize MFCC computation to make it faster and more energy-efficient for deeply embedded systems by cutting down processing time, power consumption, and memory usage. We analyzed the MFCC pipeline and introduced tailored optimizations for embedded devices while maintaining KWS accuracy. Key improvements include switching from dynamic to static memory allocation, simplifying Mel-scaling calculations, and replacing floating-point operations with fixed-point arithmetic. Our experiments achieved a 6× speedup and nearly 7× memory savings, all without loss of accuracy. These advancements make KWS applications more practical and efficient for resource-constrained embedded systems.